Reconstruct the query fan out behind a search, and find which parts of it you answer
A prompt that predicts the narrower searches an AI answer runs underneath your query, then marks which of them your site currently answers and which it ignores.
- Works in
- ChatGPT, Claude
- You need
- One query you want to be quoted on · What you sell · A list of your page titles or a sitemap, optional
- Written for
- query fan out
Scored by our own engine
This page, run through the audit we sell. Measured 8 August 2026.
A search assistant does not answer the question you typed. It writes several narrower questions, runs them at once, and assembles an answer from whatever those return. This prompt reconstructs that set for a query you care about, then marks which parts of it your site can currently be the source for.
Why the reconstruction is worth doing
Ranking data cannot explain modern citations. Pages get quoted in AI answers for queries they do not rank for at all, and the usual explanation, that the system is behaving unpredictably, is wrong. It is behaving predictably in response to questions you never saw.
Once you accept that, the planning question changes. It stops being “how do I rank for this phrase” and becomes “which of the questions underneath this phrase am I genuinely the best source for”. Those are different lists and the second one is usually shorter, cheaper and more winnable.
We wrote up the evidence behind this, including where cited pages actually rank, in how to rank in Google AI Overviews.
The column people ignore
Step 2 asks the model to name the natural source type for each sub question: a vendor, a reviewer, a community thread, documentation, research, or a video.
That column tells you where not to spend. If the natural source for “what do people regret about this” is a forum thread, a page on your own site titled the same thing is not going to take it, because a reader and a retrieval system both know a vendor is the wrong voice for that question. Publishing it anyway is a week spent on a row that was never available.
The rows where a vendor genuinely is the right source, which are typically the process, the constraints and the specifications, are where a company site wins by default and where most sites are thin.
Partial beats Missing
The output separates pages that touch a sub question from pages that answer it. The Partial rows are the ones to do first and the ones everybody skips, because a new page feels like progress and an edit does not.
An existing page already carries whatever authority it has accumulated. Adding a heading phrased as the question and a direct answer underneath it converts that page into a candidate source for one more question, at a cost of twenty minutes. A new page starts from nothing.
What it will get wrong
The decomposition is generated from what the model knows about the topic, so on subjects that changed in the last year it will miss the newest angle entirely, and the freshness column is the least reliable one in the table. It is also biased toward the way questions are asked in writing, which under represents the blunter phrasing people use when speaking to an assistant.
Both are worth checking with one real query. Ask an assistant your original question, read what it chose to cover, and compare. If the reconstruction covered most of it, the rest of the table is trustworthy for your topic. If it did not, use the structure and write the rows yourself.
For the other half of the exercise, testing what assistants say about you rather than what they ask, run the AI visibility audit prompt.
You are modelling how a search assistant decomposes a question. I will give
you one query. Produce the set of narrower searches a system would plausibly
run underneath it before writing an answer, and tell me which ones my site
is positioned to be the source for.
Step 1. Decompose.
Produce 12 to 20 sub questions the original query implies but does not state.
Cover these angles, and skip any that genuinely do not apply rather than
padding:
- Definition and scope: what the thing is, what counts and what does not.
- Selection criteria: what somebody comparing options would weigh.
- Constraints: price, time, skill, size, region, compatibility.
- Process: how it is actually done, in order.
- Failure modes: what goes wrong, what people regret, what does not work.
- Alternatives: what is chosen instead and by whom.
- Evidence: what number, study or benchmark somebody would want cited.
- Freshness: what part of the answer changed recently enough to matter.
Step 2. Classify each sub question.
- Answer shape: definition, list, comparison, number, procedure, or opinion.
- Who would be the natural source: a vendor, an independent reviewer, a
community thread, a documentation page, a research publisher, or a video.
- Whether the answer is stable or changes within a year.
Step 3. Score my coverage.
Using the pages I give you below, and only those, mark each sub question:
- Covered: a page of mine answers this directly and could be quoted for it.
- Partial: a page touches it but the answer is buried or implied.
- Missing: nothing of mine addresses it.
If I have given you no pages, write "not assessed" in this column and say
so once at the top rather than guessing at what I publish.
Output a markdown table:
Sub question | Answer shape | Natural source | Stable or changing | My coverage
Then give me, in this order:
1. The three Missing rows where I have the strongest legitimate claim to be
the source, given what I sell. Say why for each.
2. Any Partial row that could become Covered by editing an existing page
rather than writing a new one. Name the page. These are cheaper than new
pages and they are the ones people skip.
3. Any sub question where the natural source is clearly not a vendor site.
Say so plainly. If a community thread or an independent reviewer is the
right source, me publishing a page about it is unlikely to win it.
4. One sub question I would not have thought of, and why it belongs.
Constraints:
- Do not invent search volume, difficulty or traffic figures. You cannot see
them. Reason about intent and source type instead.
- Do not produce sub questions that differ only by phrasing. If two would be
answered by the same paragraph, they are one row.
- Do not claim to know the actual searches Google or any assistant runs.
This is a reconstruction from the intent, and you should say so once.
- Do not recommend a separate page for every row. Say which rows belong
together on one page.
- Do not ask me anything first. State assumptions, label them, produce the
full output, and put questions at the end.
My query: [THE QUERY]
What I sell: [ONE SENTENCE ON THE BUSINESS]
My pages: [PASTE PAGE TITLES AND URLS, OR WRITE "none"]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[THE QUERY]- One query, phrased the way a person would type or speak it. Broad commercial questions fan out furthest and are the most useful to run this on. A query that is already narrow, like a single product name, produces a short and obvious list because there is little left to decompose.
[ONE SENTENCE ON THE BUSINESS]- What you actually sell, so step 3 can tell the difference between a sub question you have a legitimate claim to and one you would be writing about because it appeared in a list. Without it every row looks like an opportunity, which is how a fan out exercise turns into forty pages nobody needed.
[PASTE PAGE TITLES AND URLS, OR WRITE "none"]- Titles and URLs are enough, and a sitemap pasted in works well. This is what makes the coverage column real rather than decorative. Write "none" and you still get the decomposition, which is the valuable half, but the model will not pretend to know what you publish.
How to run it
- 01Pick a query you want to be the answer to
Choose one you would genuinely be a good source for, not your highest volume term. The output is a list of sub questions, and the exercise only pays off where you have something real to say about several of them.
- 02Paste your page titles in
Copy the titles and URLs from your sitemap. Two minutes of work turns the coverage column from a guess into an audit, and the Partial rows it surfaces are the cheapest content work available: an edit to a page that already exists and already has whatever authority it has.
- 03Verify a few rows against a real answer
Ask the same original query in an assistant and read which sources it names and what it chose to talk about. The reconstruction is a plausible model, not a readout of anything, and comparing it to one real answer tells you quickly whether it is close for your topic.
- 04Do the Partial rows before the Missing ones
Editing an existing page to answer a sub question directly is faster than a new page and it compounds on authority the page already has. The instinct is to start with the gaps. The gaps are the slower half.
- 05Check that the new passages can be lifted
Each sub question should be answerable from your page in a passage that stands alone, roughly forty to sixty words, under a heading phrased as the question. Run the page through the AI content readiness check to confirm a clean chunk can actually be extracted rather than assuming the structure did it.
Questions people ask
What is query fan out?
The technique where a search system turns one question into several narrower questions, runs them at the same time, and composes an answer from what all of them return. The searcher never sees the sub questions. It is the reason a page can be cited on a query it does not rank for, because it ranks for one of the questions asked underneath it.
Can this prompt show me the real fan out queries Google runs?
No, and it says so in its own output. The actual sub queries are internal and not published. What this produces is a reconstruction from the intent of the original query, which is useful for planning coverage and is not a readout of anything. Treat any tool claiming to show you the literal queries with the same scepticism.
How is this different from ordinary keyword research?
Keyword research finds the phrasings people type. Fan out reconstruction finds the questions a system asks on their behalf, most of which nobody types. The two lists overlap and are not the same, and the second one explains citations that ranking data cannot.
Should I build a page for every sub question?
No, and the prompt is instructed to group them for that reason. Several sub questions usually belong in one page as separate sections with their own headings. Publishing a near identical page per row is the doorway pattern, it is a known penalty risk, and it is the most likely way this exercise does damage.
What do I do with sub questions where a vendor is the wrong source?
Accept them and spend the effort elsewhere. Some questions are naturally answered by a community thread, an independent review or a video, and a page on your own domain will not displace that. Knowing which rows those are is worth as much as knowing which ones to write, because it is the half that stops you spending a month on an unwinnable one.
Unsubscribe in one click. We never pass your address on.